Source-Free Domain Adaptation for Question Answering with Masked Self-training

TL;DR

This work tackles source-free unsupervised domain adaptation for extractive QA by proposing MDAQA, a masking-based self-training framework that preserves source-domain knowledge while adapting to a target domain without access to source data. A bottleneck mask $M$ between the encoder and QA head selects domain-relevant features, with $M$ learned on the source and refined during target adaptation through freeze/adjustment of kernel connections and gradient regularization. The method employs pseudo-label self-training on unlabeled target data across multiple rounds, using a threshold $\\alpha$ to control pseudo-label quality, and demonstrates consistent improvements over strong baselines across multiple QA datasets and backbones. The results indicate strong robustness and practicality for privacy-preserving domain adaptation, with notable gains even when target data are extremely limited. The approach also provides insight into domain discrepancy visualization and feature-map behavior, supporting its effectiveness in preserving and transferring QA knowledge under domain shifts.

Abstract

Most previous unsupervised domain adaptation (UDA) methods for question answering(QA) require access to source domain data while fine-tuning the model for the target domain. Source domain data may, however, contain sensitive information and may be restricted. In this study, we investigate a more challenging setting, source-free UDA, in which we have only the pretrained source model and target domain data, without access to source domain data. We propose a novel self-training approach to QA models that integrates a unique mask module for domain adaptation. The mask is auto-adjusted to extract key domain knowledge while trained on the source domain. To maintain previously learned domain knowledge, certain mask weights are frozen during adaptation, while other weights are adjusted to mitigate domain shifts with pseudo-labeled samples generated in the target domain. %As part of the self-training process, we generate pseudo-labeled samples in the target domain based on models trained in the source domain. Our empirical results on four benchmark datasets suggest that our approach significantly enhances the performance of pretrained QA models on the target domain, and even outperforms models that have access to the source data during adaptation.

Figures (8)

• Figure 1: Illustration of the proposed MDAQA architecture. The left part is the overall architecture, we insert a mask module between the QA encoder and the answer prediction layer. The QA encoder can be any common language model. Here we use ROBERTA. The right part is the detailed architecture of the mask module.
• Figure 2: A concise illustration of the weight adjustment associated with the mask module during source training and adaptation. All weights are free to change when the model is trained on the source domain. The red circles are kernels with mask values close to 1; thus, their output is passed to the answer prediction layer. During adaptation, the links marked in red experience minimal weight adjustments, in contrast to the remaining weights, which continue to adjust dynamically. If the mask value of some kernel changes from 0 to 1, denoted by the yellow circle, the output of that kernel will also be passed to the final layer.
• Figure 3: Impact of the number of available unlabeled target domain samples. We show the performance of MDAQA and baselines measured by EM and F1 when adapted to HotpotQA or NQ. We use logarithmic scaling for the x-axis because the performance change of the algorithms is more pronounced at low data volumes
• Figure 4: Performance Evolution Across Domains Over Five Rounds of self-training
• Figure 5: Influence of threshold $\alpha$ on adaptation performance for different datasets.